Privacy policy
Last updated: November 7, 2026. Applies to Hildegard, the app (version 0.1.x), and to Hildegard’s skills, wherever they are listed.
The short version
- Hildegard runs on your own computers. What you make, your prompts, your voices and your settings stay there, unless you turn on a service that sends them elsewhere: a rented GPU, a link or tunnel, an AI app, or a model address you add. The table below lists each one.
- The skills collect nothing and send nothing to us. They are text files and a few small scripts that an AI app reads and runs for you. Some of those scripts rely on programs you install yourself, and installing or first running those downloads files from other companies’ servers, only when you run that tool. Each one is in the table below.
- The app has no analytics, no tracking and no account with the people who make it. It sends nothing to Donovan Baney or to any server he runs.
- The app and the skills’ tools reach other companies’ servers only for the jobs in the table below. Some happen by default (checking for updates, installing, downloading a model you chose). The rest happen only when you turn them on.
The skills
Hildegard’s skills are instructions: Markdown files, a few small scripts and some page templates. They run inside the AI app you put them in, under that app’s rules and its own privacy policy. Two of their scripts can open a network connection, both listed in the table below: motion.py, for a font it has not cached yet, and cutout.py, whose rembg library downloads its model the first time. Nothing in them sends anything to Donovan Baney. The page templates load nothing from outside: no web fonts, scripts, stylesheets or images from another site. Listing the skills in a store (Claude’s directory, OpenAI’s plugin directory, Grok Bot) does not change any of that. What an AI app does with your conversation, and with any web search or page it opens for you because a skill asked it to, is between you and that app and its own policy.
What the skills do ask for is the programs their helper scripts and instructions use. Those are not included. If you or your AI app install or first run one, that program downloads from its own source. The last rows of the table below list every one the skills name with an install command or a download, and when it happens.
The app
Hildegard is free, open-source software (AGPL-3.0-only). You install it, it runs a small server on your computer that only you can reach (the loopback address, with Host and Origin checks on every request), and it does its work on your own graphics cards. Your films, stills, voices, 3D files, projects and settings live in a data folder on your computer. Passwords you set for people you invite are stored as salted scrypt hashes; keys and tokens (Hugging Face, Runpod, Figma, tunnel credentials, the keys other devices use) go to your system’s keychain, not to files or logs.
Every place the app can connect to
This list comes from reading the code. If a version of Hildegard connects somewhere not listed here, that is a bug: tell us.
| When | Where | What is sent, and why |
|---|---|---|
| Default, every six hours while the app runs | github.com (baney75/hildegard releases) | An ordinary web request for latest.json, to see whether a newer version exists. GitHub sees your IP address, as any website does. Automatic installing of an update waits until nothing is rendering and can be turned off in Settings; every update file is checked against Hildegard’s signing key first. The desktop app’s updater asks the same address. |
| Default, when you install or update | github.com (the installer, uv, ComfyUI), pypi.org and files.pythonhosted.org (Python packages, through uv), PyTorch’s package index (the GPU build of PyTorch), and the site that gives uv a Python if you have none | Download requests only. Hildegard pins each file by version and SHA-256 and refuses one that doesn’t match. |
| When you install a model | huggingface.co | A download of the model files you chose, after you accept the model’s license. If you signed in to Hugging Face, your token goes with the request so gated models download. Signing in uses Hugging Face’s own page; Hugging Face reads a small public file from this site (dbaney.com/hildegard/oauth/hf-client.json) to show the app’s name, so this site’s server may log that request. |
| Only if you turn on a link for your devices | Tailscale (tailscale serve, run by the Tailscale program on your computer); your other devices on your tailnet | Hildegard asks Tailscale to give your studio a private address. Traffic then travels over your Tailscale network, under Tailscale’s privacy policy. Sharing a GPU between your own devices (“choir”) uses the same network; each device holds a key you can revoke. |
| Only if you make your studio public (Funnel, or a Cloudflare tunnel) | Tailscale Funnel, or Cloudflare (cloudflared, dash.cloudflare.com for sign-in) | Your Hildegard’s sign-in page becomes reachable on the public internet so apps such as claude.ai, ChatGPT and Grok can reach it. Requests pass through that company’s network. Hildegard downloads cloudflared (from github.com) only when you ask, and keeps its credentials in your keychain. |
| Only if you connect an AI app (Claude, ChatGPT, Grok, Grok Bot, Codex, others) | That app’s servers, which call your Hildegard | The app sends requests to your Hildegard and receives what your Hildegard returns: the results of tools it calls, and links to videos or stills it made, which open for 24 hours without signing in. Nothing leaves unless the app asks and you signed it in. You sign in on Hildegard’s own page, and you can revoke the app in Settings. What the AI company does with the conversation is covered by its policy. |
| Only if you rent a GPU | rest.runpod.io, then Runpod’s proxy to the GPU; github.com (the pod fetches the installer) | Your Runpod API key (from your keychain), the pod you ask for, and, while it is joined to your devices, the prompts and videos for jobs you send to it. Runpod’s secure cloud runs in its own data centers; community cloud runs on other people’s machines. Hugging Face Jobs and Google Colab are described in the app but not available yet. |
| Only if you link Figma | api.figma.com | Your Figma token and the file links you ask Hildegard to read. |
| Only if you import a web page | The address you give | Hildegard loads that public page in a headless browser on your computer. It refuses addresses on your own computer or network. |
| Only if you use a language model at an address you add | The address you add (a model on your own network, or an API-key service) | The prompts for that model, sent to the address you chose. |
| Only if an AI app drafts a bug report | api.github.com; github.com | The words of the report’s title go to GitHub’s public issue search to look for a duplicate, after private details are taken out. A report is filed only after you have read the exact text and said yes, and only through your own GitHub sign-in on your computer. A filed issue is public on GitHub. |
Only when a skill’s helper cutout.py (removing a background) runs | pypi.org and files.pythonhosted.org, if rembg and onnxruntime are not installed yet (the script tells you to run pip install rembg onnxruntime); then github.com (github.com/danielgatis/rembg/releases, which sends the file on from GitHub’s download servers) | Download requests only. The first run of rembg downloads its model (birefnet-general by default, or the isnet-general-use or u2net one you pick) once and keeps it in a .u2net folder in your home folder. Your picture is not uploaded: it is processed on your computer. The sites see your IP address. |
| Only when your AI app runs another install a skill suggests | pypi.org and files.pythonhosted.org (through pip); registry.npmjs.org (through npm) | Download requests only. The skills name pip install scikit-image (solid-model checks in hildegard-3d), pip install bpy (Blender as a Python module), pip install pronouncing (syllable counts for songs), pip install librosa soundfile (beat and energy analysis for a music video), pip install av numpy pillow (the motion and music skills’ scripts), pip install playwright then python3 -m playwright install chromium (an optional browser for the motion skill, which also downloads Chromium from Microsoft’s Playwright servers), and the @fontsource font packages on npm, which an image task fetches so text renders in the right typeface. The helper scripts also import numpy, Pillow and scipy, which come from your own Python. Each runs only if you or your AI app, with your permission, run the command. |
Only when the motion skill’s motion.py draws text in a font it has not cached yet | raw.githubusercontent.com (Google’s font repository, google/fonts, at one pinned commit) | A download request for that one font file, which is checked against its SHA-256 and kept in a cache folder, so it is never fetched again. HILDEGARD_MOTION_OFFLINE=1 turns this off. The site sees your IP address. |
Only when the image skill’s cutout.py removes a background for the first time | github.com (the rembg project’s model release) | rembg, a library you install yourself, downloads its model file once and keeps it in ~/.u2net. The site sees your IP address. No other helper script opens a connection. |
| Only if you open a link in a skill’s reference page | The site you open, for example a model’s license at stability.ai or huggingface.co | Your browser’s visit, as with any link. The skills only cite these pages; nothing in them fetches one. |
| Only if you follow the setup or cloud skill | github.com (the Hildegard installer); console.runpod.io (a link you open to make an API key) | The setup skill (in OpenAI’s package and Grok Bot’s zips, not in Claude’s listing) tells an AI app to run the same installer command as the install row above. The cloud skill shows you a Runpod page to open yourself; nothing is sent from the skill. |
The skills also name other programs without an install command: FFmpeg, ImageMagick, OpenSCAD, Blender, and model tools such as whisperx, allin1, demucs and SAM 2. They are not part of this package. If you choose to install one, you get it from its own site, and some of them download model files on first use under their own privacy terms. The skills never install or run them on their own. Skills that tell an AI app to search the web or open a product page (research, the 3D-printing measurements) use that app’s own web tool, not code in the skills.
Voice cloning is consent-gated and local: the voice samples, the consent recording and the cloned voices stay on your computer, and Hildegard never lets an AI app start a clone by itself.
What Donovan Baney receives
Nothing from the app or the skills. There is no telemetry, no crash reporter and no account. If you open an issue or a pull request on GitHub, what you write there is public and is held by GitHub. If you email support@dbaney.com, the message and your address are kept to answer you.
This website
dbaney.com/hildegard is a static page: no cookies, no advertising and no analytics scripts. The install page remembers which operating system tab you picked, in your own browser’s local storage, and sends it nowhere. It is served by Cloudflare, which logs a sample of about one request in ten (the address, the page asked for, the time and the browser it named) and keeps those logs for at most 7 days. They are used only to keep the site running.
Your choices
- Turn off automatic updates in Settings, or run the app offline after install; every render works offline once the models are downloaded.
- Revoke any app, key or device in Settings › Sharing, or with
hildegard host keys revokeandhildegard choir revoke. - Turn off a link or tunnel with
hildegard host offorhildegard tunnel off. - Delete your data by deleting Hildegard’s data folder and its entries in your keychain.
Changes and contact
Changes to this page are made in the open, in the project’s repository, and the date at the top changes with them. Questions or concerns: github.com/baney75/hildegard/issues, or support@dbaney.com.